Advancing the Discovery of Phage-Host Interactions and Disease Classification from Metagenomic Profiles Using Deep Learning

Microbiomes, the communities of microorganisms, are essential for sustaining ecosystem functions in diverse environments, including the human gut. Bacteriophages (phages) interact dynamically with their prokaryotic hosts and play a crucial role in shaping the structure and function of microbial communities. Previous approaches for inferring phage-host interactions from metagenomic data (e.g., assembly-based methods) are constrained by high computational demands, limited sensitivity, and the inability to accurately capture ecological relationships. To address these issues, we developed phiNODE (<underline>p</underline>hage-<underline>h</underline>ost interaction predictor using<underline>N</underline>eural<underline>O</underline>rdinary<underline>D</underline>ifferential<underline>E</underline>quations), a deep learning method for predicting phage-host interactions directly from metagenomic profiles. We first validated phiNODE using synthetic datasets generated by ecological models and found that it outperformed both alternative deep-learning and co-abundance-based methods in inferring phage-host interactions. We then applied phiNODE to a large-scale metagenomic dataset comprising 7,016 stool samples from healthy individuals and identified 90% more genus-level phage-host interactions than traditional assembly-based methods. Finally, we demonstrated that the latent representations learned by phiNODE served as more powerful features than taxonomic abundance-based features for disease classifications. In summary, phiNODE offers a novel framework for inferring phage-host interactions from shotgun metagenomic data, representing an innovative approach for advancing microbiome research and clinical applications.